Generate_To_Adapt

Implementation of "Generate To Adapt: Aligning Domains using Generative Adversarial Networks" in PyTorch

Datasets:

Please download the dataset from http://www.cs.umd.edu/~yogesh/datasets/digits.zip and extract it. This folder contains the dataset in the same format as need by our code.

Cifar10 & STL10 dataset: https://drive.google.com/open?id=1kWpIt9W0XtU0p8V8om3MqOijurwmuLoU

Training:

Dataset Selection

1. svhn->mnist
2. mnist->svhn
3. cifar10->stl10
4. stl10->cifar10

Train

Let us train the Lenet model for SVHN->MNIST Domain adaptation. Obtain the baseline numbers by running

python main.py --dataselect 1 --dataroot [path to the dataset] --method sourceonly

To train our method(GTA), run

python main.py --dataselect 1 --dataroot [path to the dataset] --method GTA

This code trains and stores the trained models in result folder. Current checkpoint and the model that gives best performance on the validation set are stored.

Evaluation:

To evaluate the trained models on the target domain (MNIST), run

python eval.py --dataselect 1 --dataroot [path to the dataset] --method GTA --model_best False

Citation:

If you use this code for your research, please cite

@article{Gen2Adapt,
    author    = {Swami Sankaranarayanan and
           Yogesh Balaji and
           Carlos D. Castillo and
           Rama Chellappa},
    title     = {Generate To Adapt: Aligning Domains using Generative Adversarial Networks},
    journal   = {CoRR},
    volume    = {abs/1704.01705},
    year      = {2017},
    url       = {http://arxiv.org/abs/1704.01705},
}